Merge pull request #19764 from hrydgard/even-more-depth-raster-stuff

Enable depth raster in all backends

Common/Math/CrossSIMD.h

@@ -172,7 +172,7 @@ struct Vec4U16 {
 	Vec4U16 operator ^(Vec4U16 other) const { return Vec4U16{ _mm_xor_si128(v, other.v) }; }
 
 	Vec4U16 Max(Vec4U16 other) const { return Vec4U16{ _mm_max_epu16_SSE2(v, other.v) }; }
-	Vec4U16 Min(Vec4U16 other) const { return Vec4U16{ _mm_max_epu16_SSE2(v, other.v) }; }
+	Vec4U16 Min(Vec4U16 other) const { return Vec4U16{ _mm_min_epu16_SSE2(v, other.v) }; }
 	Vec4U16 CompareLT(Vec4U16 other) { return Vec4U16{ _mm_cmplt_epu16(v, other.v) }; }
 };
 

GPU/Common/DepthRaster.cpp

@@ -199,7 +199,7 @@ TriangleResult DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y
 			switch (compareMode) {
 			case ZCompareMode::Greater:
 				// To implement the greater/greater-than comparison, we can combine mask and max.
-				// It might be better to do the math in float space on x86 due to SSE2 deficiencies.
+				// Unfortunately there's no unsigned max on SSE2, it's synthesized by xoring 0x8000 on input and output.
 				// We use AndNot to zero out Z results, before doing Max with the buffer.
 				AndNot(shortZ, shortMaskInv).Max(bufferValues).Store(rowPtr + x);
 				break;

GPU/D3D11/DrawEngineD3D11.cpp

@@ -335,6 +335,9 @@ void DrawEngineD3D11::Flush() {
 				context_->Draw(vertexCount, 0);
 			}
 		}
+		if (useDepthRaster_) {
+			DepthRasterTransform(prim, dec_, dec_->VertexType(), vertexCount);
+		}
 	} else {
 		PROFILE_THIS_SCOPE("soft");
 		VertexDecoder *swDec = dec_;
@@ -388,6 +391,13 @@ void DrawEngineD3D11::Flush() {
 			UpdateCachedViewportState(vpAndScissor);
 		}
 
+		// At this point, rect and line primitives are still preserved as such. So, it's the best time to do software depth raster.
+		// We could piggyback on the viewport transform below, but it gets complicated since it's different per-backend. Which we really
+		// should clean up one day...
+		if (useDepthRaster_) {
+			DepthRasterPredecoded(prim, decoded_, numDecodedVerts_, swDec, vertexCount);
+		}
+
 		SoftwareTransform swTransform(params);
 
 		const Lin::Vec3 trans(gstate_c.vpXOffset, -gstate_c.vpYOffset, gstate_c.vpZOffset * 0.5f + 0.5f);

GPU/Directx9/DrawEngineDX9.cpp

@@ -292,6 +292,9 @@ void DrawEngineDX9::Flush() {
 				}
 			}
 		}
+		if (useDepthRaster_) {
+			DepthRasterTransform(prim, dec_, dec_->VertexType(), vertexCount);
+		}
 	} else {
 		VertexDecoder *swDec = dec_;
 		if (swDec->nweights != 0) {
@@ -344,6 +347,13 @@ void DrawEngineDX9::Flush() {
 			UpdateCachedViewportState(vpAndScissor);
 		}
 
+		// At this point, rect and line primitives are still preserved as such. So, it's the best time to do software depth raster.
+		// We could piggyback on the viewport transform below, but it gets complicated since it's different per-backend. Which we really
+		// should clean up one day...
+		if (useDepthRaster_) {
+			DepthRasterPredecoded(prim, decoded_, numDecodedVerts_, swDec, vertexCount);
+		}
+
 		int maxIndex = numDecodedVerts_;
 		SoftwareTransform swTransform(params);
 

GPU/GLES/DrawEngineGLES.cpp

@@ -315,6 +315,9 @@ void DrawEngineGLES::Flush() {
 				inputLayout, vertexBuffer, vertexBufferOffset,
 				glprim[prim], 0, vertexCount);
 		}
+		if (useDepthRaster_) {
+			DepthRasterTransform(prim, dec_, dec_->VertexType(), vertexCount);
+		}
 	} else {
 		PROFILE_THIS_SCOPE("soft");
 		VertexDecoder *swDec = dec_;
@@ -371,6 +374,13 @@ void DrawEngineGLES::Flush() {
 			}
 		}
 
+		// At this point, rect and line primitives are still preserved as such. So, it's the best time to do software depth raster.
+		// We could piggyback on the viewport transform below, but it gets complicated since it's different per-backend. Which we really
+		// should clean up one day...
+		if (useDepthRaster_) {
+			DepthRasterPredecoded(prim, decoded_, numDecodedVerts_, swDec, vertexCount);
+		}
+
 		SoftwareTransform swTransform(params);
 
 		const Lin::Vec3 trans(gstate_c.vpXOffset, gstate_c.vpYOffset, gstate_c.vpZOffset);